Brain Tumor Image Segmentation Based On Convolutional Neural Network

Glioma,known as a common primary brain malignant tumor,seriously threatens human health.Based on MRI technology,doctors analyze and diagnose patient’s condition by constructing brain tumor images and performing accurate and automatic segmentation of brain tumor images.Traditional tumor image segmentation is mainly based on threshold,region,edge and other methods,and the calculation speed is fast.However,the accuracy of image segmentation is low due to the dependence on userspecified parameters and image preprocessing.While,the network segmentation model relied on deep learning can produce better results in image segmentation.However,the problems of blurred boundaries and low segmentation accuracy in two-dimensional brain tumor image segmentation methods,as well as shortcomings of large parameters and calculations in three-dimensional brain tumor image segmentation methods have been in existence.Complex and other issues,therefore,this paper proposes a two-dimensional attention Unet network and a three-dimensional lightweight Vnet network for existing problems.This paper mainly includes the following aspects:First of all,in response to the current two-dimensional MRI tumor image,the tumor boundary is blurred and the segmentation is difficult,and the two-dimensional attention Unet network is proposed.Based on the Unet network structure,the Inception upsampling\down-sampling module and feature fusion attention module are added.Because the Inception up-sampling and down-sampling module uses the idea of Inception structure,it can capture spatial information at different resolutions,and perform upsampling and down-sampling and fusion of feature maps of different resolutions,so that the point-like,Small tumors that are discontinuous and indistinguishable by the naked eye are presented.The feature fusion attention module learns the local and lowdimensional features of the image through skip links,and provides richer and more detailed information for the network.The ablation experiment proved the effectiveness of each module and achieved accurate segmentation of brain tumor MRI images.Moreover,in view of the large parameters and complex calculations of the current three-dimensional brain tumor image segmentation algorithm,this paper proposes a three-dimensional lightweight V-shaped convolutional neural network,which uses threedimensional expansion separable convolution to extract different depths Spatial structure information,ensures the integrity of information and avoids information loss caused by continuous pooling.In this paper,the three-dimensional lightweight Vnet segmentation network attempts to extract more image information as well as obtain better results in segmentation without increasing model parameters.Finally,this article uses the Brats18 dataset containing high-grade gliomas and lowgrade gliomas for training and testing.The three-dimensional data set is divided into twodimensional images and three-dimensional images,which are input to the twodimensional attention Unet network and the three-dimensional lightweight Vnet,respectively.The internet.Through the experiment of segmenting brain tumor images with different models in the data set,the effectiveness and adaptability of the algorithm proposed in this paper are verified.The segmentation results are evaluated by Dice segmentation coefficient and Sensitivity.Compared with Vnet3 d,RSANet3d,Unet3+ and other networks,the two networks proposed in this paper improve the accuracy of MRI brain tumor image segmentation.